Refusal-text leakage
- Id
- refusal-text-leakage
- Status
- Active
- Severity
- high
- Detection
- deterministic
- Evidence grade
- corroborated
- Languages
- en
- Added
- 2026-08-15
- Updated
- 2026-08-17
Currently signals low-effort writing.
What it is
The published text contains the model refusing the prompt. A product title that declines to write a product title. A paragraph that apologises and cites a use policy. The artifact and the boilerplate shipped together because nobody read the output before it went live. This evidences a failed review, and it says nothing about who wrote the sentences around it.
Why it reads as machine-made
Refusal boilerplate has one origin. A model writes it to a prompter about a prompt. Amazon listings carried titles of exactly this shape in January 2024, on furniture and other ordinary goods. NewsGuard names error messages and other language specific to chatbot responses among the markers it uses to find AI content farms, a count it put at 3,749 sites in June 2026. Wikipedia deletes pages on the same sign under speedy-deletion criterion G15, which files it as communication intended for the user.
Specimens
Product title: Sorry, I cannot fulfill this request. Ergonomic Office Chair, Mesh Back, Adjustable Lumbar Support.
Ergonomic Office Chair, Mesh Back. Lumbar support adjusts across 7 cm, the gas lift is rated to 120 kg, and the box ships flat at 14 kg.
The repair replaces the leaked line with the three measurements a buyer filters on. Figures in this repair are invented for the specimen.
Council business continued as normal this week. I cannot generate a response to that request. Residents will see the effects of the decision over the coming months, and officials say the new arrangements are designed to serve the community better. A timetable for the next stage has yet to be confirmed, though members have asked that the concerns raised during consultation be carried forward into whatever follows. The authority remains committed to keeping residents informed as the arrangements bed in.
The council approved the budget on Tuesday by 31 votes to 12. The new parking charges start on 1 September and raise the day rate in the two central zones from 2 euro to 3 euro.
A body-copy sighting rather than a title. The repair fills the hole the leaked sentence was sitting in. Figures in this repair are invented for the specimen.
How it is detected
- Pattern
- \bI(?:'|’)?m sorry[,.]?\s+but\s+I\s+(?:cannot|can(?:'|’)?t|am unable to)\b[^.!?\n]{0,60}\b(?:request|prompt|content)\b|\bI\s+(?:cannot|can(?:'|’)?t)\s+fulfil{1,2}\s+(?:this|that|your)\s+request\b|\b(?:goes|would go)\s+against\s+(?:my|the|OpenAI(?:'|’)?s?|Anthropic(?:'|’)?s?)\s+(?:use|usage|content)\s+polic(?:y|ies)\b|\bI\s+(?:cannot|can(?:'|’)?t)\s+generate\s+a\s+response\s+to\s+(?:this|that|your)\b
- Flags
- i
- Scope
- sentence
Who writes this way legitimately
Reporters, researchers and editors quote refusal text deliberately. This entry does it twice in its own specimens. Anyone covering model behaviour, teaching prompt design or filing a bug report reproduces the string exactly, and a lint that cannot see quotation marks flags all of them. Exclude quoted spans, code samples and transcribed screenshots before a hit counts. What is left is narrow: refusal boilerplate in running copy, with no quotation frame and nothing anywhere near it about assistants. Support macros that decline a request are a separate case, since a human wrote and approved the macro.
Model attribution
Vendor-neutral in principle. The Amazon titles named an OpenAI use policy, which records which product was cheapest to run in bulk in early 2024. It is not a property of one model family. Every assistant emits refusal text in some shape, and the wording moves between releases.
Platform notes
- amazon
- The January 2024 sighting was product titles on furniture and other goods. Amazon describes its fake-review and abuse detection as behavioural and graph-based, so a listing like this is caught by a reader, not by a style rule.
- wikipedia
- Criterion G15 allows speedy deletion where a page shows communication intended for the user. Refusal and safety boilerplate is the clearest member of that class, and it is the one editors report as unmistakable on sight.
Status history
| Date | Status | Rationale |
|---|---|---|
| 2026-08-15 | Active | Active on three independent records: the Amazon product titles reported in January 2024, NewsGuard listing chatbot error messages among the signals it uses to identify AI content farms, and Wikipedia deleting pages under G15 for the same sign. Pattern verified on 2026-08-15 against both specimens, against both repairs, and against three ordinary sentences including one about a refund request that was never fulfilled. |
Sources
- 01Futurism: Amazon products with AI refusal text as titlespressaccessed 2026-08-14
- 02NewsGuard AI Tracking Center and the UAIN frameworkpressaccessed 2026-08-14
- 03Wikipedia: Criteria for speedy deletion, criterion G15communityaccessed 2026-08-14
CC BY 4.0 / The AI Tells Index, feedsquad.com/ai-tells